Fix an intermittent hang during GRPO generation with a PEFT/LoRA model and colocated vLLM tensor parallelism…
Fix an intermittent hang during GRPO generation with a PEFT/LoRA model and colocated vLLM tensor parallelism…: a task in HF ML Tasksmith (Harbor dataset). Ensure the training workers have finished the relevant shared communication before colocated tensor-parallel inference begins. Keep generation…
The task
Ensure the training workers have finished the relevant shared communication before colocated tensor-parallel inference begins. Keep generation results and per-rank prompt/completion slicing unchanged. Ordinary models and single-device tensor parallelism should keep their existing behavior without extra…
Part of FineEnvs/HF_ML_Tasksmith.